Drift va qayta o'qitish: taksichini qachon o'qitish kerak

import pandas as pd
show(pd.DataFrame([
    {"Turi": "Ma'lumot drifti", "Nima o'zgardi": "Kiruvchi ma'lumot taqsimoti",
     "Yorliqsiz seziladimi": "**Ha**", "Qayta o'qitish yordam beradimi": "Odatda ha"},
    {"Turi": "Tushuncha drifti", "Nima o'zgardi": "Kirish va javob orasidagi bog'liqlik",
     "Yorliqsiz seziladimi": "Qiyin", "Qayta o'qitish yordam beradimi": "Yangi yorliq kerak"},
    {"Turi": "Xususiyat buzilishi", "Nima o'zgardi": "Quvur (3, 6-bob)",
     "Yorliqsiz seziladimi": "Ha — keskin sakrash", "Qayta o'qitish yordam beradimi": "**Yo'q — tuzatish kerak**"},
]))

python3 -c "
import pandas as pd
import numpy as np

eski = pd.read_csv('../datasets/ijara.csv')
yangi = pd.read_csv('../datasets/ijara_yangi.csv')
print(f'eski: {len(eski):,} qator | yangi: {len(yangi):,} qator')
print()
USTUN = ['tavsif_uzunligi','rasmlar_soni','telefon_takrorlanishi','kunlik_elonlar','narx_mln']
print(f'{\"ustun\":22s} {\"eski ort\":>10s} {\"yangi ort\":>10s} {\"farq %\":>9s}')
print('-' * 55)
for u in USTUN:
    a, b = eski[u].mean(), yangi[u].mean()
    print(f'{u:22s} {a:>10.2f} {b:>10.2f} {(b-a)/a*100:>8.1f}%')
"

"""Taqsimot siljishini o'lchash — yorliqsiz."""
import numpy as np
import pandas as pd

# sanoatda qabul qilingan chegaralar
KICHIK, KATTA = 0.10, 0.25


def psi(asos: pd.Series, yangi: pd.Series, paqir: int = 10) -> float:
    """Population Stability Index.

    < 0.10  - siljish yo'q
    < 0.25  - o'rtacha siljish, kuzating
    >= 0.25 - katta siljish, harakat kerak
    """
    chegara = np.quantile(asos, np.linspace(0, 1, paqir + 1))
    chegara[0], chegara[-1] = -np.inf, np.inf
    a, _ = np.histogram(asos, bins=chegara)
    b, _ = np.histogram(yangi, bins=chegara)
    # nol bo'lmasin - logarifm uchun
    a = np.clip(a / max(a.sum(), 1), 1e-6, None)
    b = np.clip(b / max(b.sum(), 1), 1e-6, None)
    return float(np.sum((b - a) * np.log(b / a)))


def baho(qiymat: float) -> str:
    if qiymat < KICHIK:
        return "barqaror"
    if qiymat < KATTA:
        return "o'rtacha"
    return "KATTA SILJISH"


def hisobot(asos: pd.DataFrame, yangi: pd.DataFrame,
            ustunlar: list[str]) -> pd.DataFrame:
    qatorlar = [
        {"ustun": u, "psi": round(psi(asos[u], yangi[u]), 4),
         "baho": baho(psi(asos[u], yangi[u]))}
        for u in ustunlar
    ]
    return pd.DataFrame(qatorlar).sort_values("psi", ascending=False)

python3 -c "
import sys; sys.path.insert(0, 'src')
import pandas as pd
from quvur.drift import hisobot

eski = pd.read_csv('../datasets/ijara.csv')
yangi = pd.read_csv('../datasets/ijara_yangi.csv')
USTUN = ['tavsif_uzunligi','rasmlar_soni','telefon_takrorlanishi','kunlik_elonlar',
         'narx_mln','maydon','xonalar','egasi_tasdiqlangan']
h = hisobot(eski, yangi, USTUN)
print(h.to_string(index=False))
"

python3 -c "
import sys; sys.path.insert(0, 'src')
import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import f1_score, precision_score, recall_score
from sklearn.model_selection import train_test_split
from quvur.xususiyat import tayyorla

eski = pd.read_csv('../datasets/ijara.csv')
yangi = pd.read_csv('../datasets/ijara_yangi.csv')

x, y = tayyorla(eski), eski['broker']
Xtr, Xte, ytr, yte = train_test_split(x, y, test_size=0.25, stratify=y, random_state=42)
m = RandomForestClassifier(n_estimators=150, min_samples_leaf=3, random_state=42, n_jobs=1).fit(Xtr, ytr)

def baho(mod, X, Y):
    p = mod.predict(X)
    return f1_score(Y,p), precision_score(Y,p), recall_score(Y,p)

a = baho(m, Xte, yte)
b = baho(m, tayyorla(yangi), yangi['broker'])
print(f'{\"davr\":14s} {\"F1\":>8s} {\"aniqlik\":>9s} {\"toliqlik\":>9s}')
print('-' * 44)
print(f'{\"eski (sinov)\":14s} {a[0]:>8.3f} {a[1]:>9.3f} {a[2]:>9.3f}')
print(f'{\"yangi\":14s} {b[0]:>8.3f} {b[1]:>9.3f} {b[2]:>9.3f}')
print()
print(f'F1 tushishi: {a[0]-b[0]:.3f}')
"

python3 -c "
import sys; sys.path.insert(0, 'src')
import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import f1_score
from sklearn.model_selection import train_test_split
from quvur.xususiyat import tayyorla

eski = pd.read_csv('../datasets/ijara.csv')
yangi = pd.read_csv('../datasets/ijara_yangi.csv')

Xy, yy = tayyorla(yangi), yangi['broker']
Xtr_y, Xte_y, ytr_y, yte_y = train_test_split(Xy, yy, test_size=0.25, stratify=yy, random_state=42)

# 1) ESKI model, yangi sinovda
xe, ye = tayyorla(eski), eski['broker']
Xtr_e, _, ytr_e, _ = train_test_split(xe, ye, test_size=0.25, stratify=ye, random_state=42)
eski_m = RandomForestClassifier(n_estimators=150, min_samples_leaf=3, random_state=42, n_jobs=1).fit(Xtr_e, ytr_e)
f_eski = f1_score(yte_y, eski_m.predict(Xte_y))

# 2) FAQAT yangi malumotda oqitilgan
yangi_m = RandomForestClassifier(n_estimators=150, min_samples_leaf=3, random_state=42, n_jobs=1).fit(Xtr_y, ytr_y)
f_yangi = f1_score(yte_y, yangi_m.predict(Xte_y))

# 3) IKKALASI birga
Xb = pd.concat([Xtr_e, Xtr_y]); yb = pd.concat([ytr_e, ytr_y])
birga_m = RandomForestClassifier(n_estimators=150, min_samples_leaf=3, random_state=42, n_jobs=1).fit(Xb, yb)
f_birga = f1_score(yte_y, birga_m.predict(Xte_y))

print('YANGI sinov toplamida F1:')
print(f'  eski model          : {f_eski:.3f}')
print(f'  faqat yangida oqitildi: {f_yangi:.3f}')
print(f'  ikkalasida oqitildi : {f_birga:.3f}')
print()
print(f'qayta oqitish foydasi: +{f_yangi-f_eski:.3f}')
"

diagramma(
    {
        "Ishlayotgan\nmodel": (0, 2.0, 2.2, 1.0, "xizmat"),
        "Monitoring\n(34-bob)": (3.0, 2.0, 2.2, 1.0, "kuzatuv"),
        "Drift\nsezildi": (6.0, 2.0, 2.0, 1.0, "kuzatuv"),
        "Sabab?\nsifat yoki drift": (8.8, 2.0, 2.4, 1.0, "qayta"),
        "Quvurni\ntuzatish": (11.9, 3.4, 2.1, 0.9, "manba"),
        "Qayta\no'qitish": (11.9, 0.7, 2.1, 0.9, "qayta"),
    },
    [
        ("Ishlayotgan\nmodel", "Monitoring\n(34-bob)"),
        ("Monitoring\n(34-bob)", "Drift\nsezildi"),
        ("Drift\nsezildi", "Sabab?\nsifat yoki drift"),
        ("Sabab?\nsifat yoki drift", "Quvurni\ntuzatish", "nosozlik"),
        ("Sabab?\nsifat yoki drift", "Qayta\no'qitish", "haqiqiy drift"),
    ],
    nom="drift_halqa.png", kenglik=15,
    izoh="Eng muhim tugun - o'rtadagi savol. Usiz drift har doim 'qayta o'qitish' deb javob oladi.",
)